报告题目：Solving Markov Decision Processes via Largest-Size Average Estimator
报告人：刘光梧 香港城市大学
报告时间：2023年3月15日(周三) 16:00-17:30
报告地点：中国科学院大学中关村校区教学楼S406
腾讯会议：873-9767-9281
内容摘要
Within an adaptive multistage sampling (AMS) framework, we propose to solve finite-horizon Markov decision processes (MDPs) by incorporating the so-called largest-size average (LSA) estimator which approximates the value function by the sample average for the action with the largest size of allocated samples. The LSA-based AMS algorithm sequentially draws samples from actions using the upper confidence bound sampling method. We show the consistency and asymptotic unbiasedness of the algorithm. In particular, we show that the bias of the proposed algorithm decays at a rate that is much faster than an existing algorithm. The superiority of the LSA-based algorithm over several alternatives is also demonstrated via numerical examples, especially when sample sizes are relatively small.
主讲人简介
Dr. Guangwu Liu is currently a professor in Department of Management Sciences, College of Business at City University of Hong Kong. His research interests include stochastic simulation and machine learning, with applications in financial engineering and risk management. He has published in journals including INFORMS Journal on Computing, Management Science, Operations Research, and Production and Operations Management. He currently serves as an associate editor for Naval Research Logistics and Operations Research.
